freebsd-src/UPDATING.64BTT
Garance A Drosehn 01bfe33958 Add a tip for people who are using database-related ports on a sparc64
system that they intend to upgrade to 64-bit time_t.

Noticed by:	Roderick van Domburg
2004-03-17 01:59:47 +00:00

369 lines
19 KiB
Plaintext

# -------+---------+---------+---------+---------+---------+---------+---------+
The FreeBSD/sparc64 port has now changed time_t from 32-bits to 64-bits.
This file explains the exact steps that users should follow to update their
sparc64 systems for this change. People running FreeBSD on other types of
hardware, such as CPU's from Intel or AMD, can ignore this file. For now,
this change is only happening for people running FreeBSD on Sparc hardware.
# -------+---------+---------+---------+---------+---------+---------+---------+
# Copyright (c) 2004 - Garance Alistair Drosehn <gad@FreeBSD.org>.
#
# All rights reserved.
#
# Redistribution, publication, translation and use, with or without
# modification, in full or in part, in any form or format of this
# document are permitted without further permission from the author.
#
# THIS DOCUMENT IS PROVIDED BY GARANCE DROSEHN ``AS IS'' AND ANY EXPRESS
# OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
# WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
# DISCLAIMED. IN NO EVENT SHALL GARANCE DROSEHN BE LIABLE FOR ANY DIRECT,
# INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
# (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
# HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
# STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
# IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
# POSSIBILITY OF SUCH DAMAGE.
#
# -------+---------+---------+---------+---------+---------+---------+---------+
# $FreeBSD$
# -------+---------+---------+---------+---------+---------+---------+---------+
If you are in too much of a hurry to read this file, then this is not the
time for you to upgrade to a 64-bit time_t. Period. Stick with a system
using 32-bit time_t until you have plenty of time to perform an upgrade.
This statement is true even if you have performed a thousand system upgrades
in the past, and you are certain that you know everything there is to know
about upgrades. This upgrade *will* take you more time than previous system
upgrades, simply because you must recompile at least some of your ports after
upgrading the base system.
Do not start this update unless you have the extra time.
As of March 10th, the official value for time_t on sparc64 has changed to
be 64-bits. If you really must build system with 32-bit time_t's, then it
would be best to stick with a snapshot of current from before March 10th.
With src snapshots after that point, there will soon be ports which assume
you have a 64-bit time_t based on the value of __FreeBSD_version.
If you are ready to upgrade, then *READ THIS ENTIRE DOCUMENT* at least
once before starting the upgrade.
# -------+---------+---------+---------+---------+---------+---------+---------+
This is a major change. This change will *not* be backwards-compatible.
Any programs which call system-routines for handling time-values will
have to be recompiled after this change is made.
Because this change is not backwards-compatible, it is important that
the following steps be used when upgrading the system. "Shortcuts" that
have worked for EVERY SINGLE UPGRADE YOU HAVE EVER DONE IN YOUR LIFE are
probably irrelevant. This change is more disruptive than most of the
changes which are normally done on freebsd.
These steps are designed to minimize the chance of you running into any
trouble. We can not guarantee that these steps will avoid all possible
problems, but if you ignore these steps you are very likely to run into
some very painful and time-consuming headaches when upgrading.
Step Pre-1: Update to a recent snapshot of -current, and first build
that as a system with 32-bit time_t. To do this, edit
the file /usr/src/sys/sparc64/include/_types.h
find the line:
typedef __int64_t __time_t; /* time()... */
and change '__int64_t' to '__int32_t'
Step Pre-2: Install that system, using whatever steps you normally
use, and make sure that installation seems to work okay.
Step Pre-3: While still running that 32-bit time_t system, it would
probably be a good idea to cvsup your ports tree, and
then upgrade portupgrade (if you use it) and upgrade any
shells that you use. Eg:
portupgrade -Rr -f ruby portupgrade
portupgrade -Rr -f bash
That way you know you have the latest versions, and you
will also know you have the most-recent distfiles on
your machine.
Step Pre-4: For sparc64 machines which need DHCP:
The 'dhclient' in the base system is known to be unreliable
on a system which is upgraded to 64-bit time_t's. It may
work for you, but it probably will not.
As of March 10th 2004, we have no fix for that.
However, the net/isc-dhcp3-client port does seem to work.
IF your machine needs DHCP, then you should probably install
that port and make sure you can get it working *before* you
make the change to use 64-bit time_t's.
Step Pre-4: For people using database-related ports:
Ports like PostgreSQL may change how they store data after
they are recompiled for 64-bTT. So, you may have to do a
full dump of your data while you still have a 32-bTT system,
and do a matching initdb/restore of your data after the port
has been recompiled on the upgraded 64-bTT system.
After you have built and installed that src-snapshot with 32-bit time_t's,
edit the file /usr/src/sys/sparc64/include/_types.h
find the line:
typedef __int32_t __time_t; /* time()... */
and change '__int32_t' back to '__int64_t'
For best results, do NOT make any other changes. Do NOT cvsup the
source tree trying to pick up any other changes. At this point you
know that you have a source tree that does work for your system, so
stick with that source tree (except for making the above 1-line
change, of course).
At one point in my testing, I did do a 'cvsup' which just happened
to pull in one bad commit that broke 'make buildworld', and a second
bad commit that broke 'make installworld'. Believe me, you REALLY
REALLY do *not* want to risk problems like that!
I am not suggesting that you have to do two whole buildworld/
installworld cycles in a single day. You could easily wait a few
days, or even a week between them. What I am suggesting is that
you should not 'cvsup' your sources inbetween the two buildworlds.
And then follow these steps to build and install the 64-bit time_t system:
cd /usr/src #- 1.
make cleanworld #- 2. or 'rm -Rf /usr/obj/usr/src/*'
make buildworld #- 3.
make buildkernel #- 4. Add KERNCONF if you usually do.
NEWSPARC_TIMETYPE=__int64_t #- 5. (Used by a safety-check done
export NEWSPARC_TIMETYPE #- 5a. by installkernel)
make installkernel #- 6. Add KERNCONF if you usually do.
mergemaster -p #- 7.
# - - A section required for installs over NFS-mounts - - #
ifconfig -a #- NFS 8a. See note below.
shutdown now #- NFS 8b. NOT 'shutdown -r now'
cd /usr/src #- NFS 8c.
sh installworld_oldk #- NFS 8d. See note below.
# - - End of this section for NFS-mounts - - #
reboot #- 9. MUST go into single-user mode
For many upgrades, it is true that you can "cheat" at this point, and
get away without actually going into single-user mode straight from
the reboot. But for this upgrade, you REALLY MUST start up straight
into single user mode. So, reboot the machine, type a space (or
anything other than 'Enter') when the boot-loader is counting down.
And then:
boot -s #- 10. (command to boot-loader)
The system will ask you if you want to use /bin/sh or some other shell.
For this upgrade, just hit enter, even if you usually prefer like some
other shell instead of /bin/sh.
fsck -p #- 11.
# - - A section required for installs over NFS-mounts - - #
PATH=/boot/kernel/bin:$PATH #- NFS 12.
# - - End of this section for NFS-mounts - - #
mount -a -t ufs #- 13.
swapon -a #- 14.
# - - A section required for installs over NFS-mounts - - #
ifconfig hme0 inet .... #- NFS 15a. See note below.
mount_nfs host:srcdir /usr/src #- NFS 15b. See note below.
mount_nfs host:objdir /usr/obj #- NFS 15c.
# - - End of this section for NFS-mounts - - #
cd /usr/src #- 16.
sh installworld_newk #- 17. Might want to add -S
mergemaster #- 18.
rm -f /var/db/dhclient.leases #- 19. If this host uses DHCP
reboot #- 20.
At this point, you should be up-and-running on a system that has 64-bit
values for time_t. You will have to rebuild anything which depends on
time_t. Later in this file is a suggested order for upgrading ports.
If you have a lot of ports which start up daemons or do other processing
at system-startup, then you might want to have this reboot also go into
single-user mode for upgrading all of the ports. In my case, I've always
done a standard reboot at this point and did not run into problems, but
then I only have 25 ports installed on my SPARC64 system.
Aside: It is slightly more reasonable to use the 'reboot' command, although
you may be more familar with using 'shutdown -r now'. The shutdown command
just turns around and executes '/sbin/reboot', and with this upgrade it is
best to avoid such redirection.
# -------+---------+--------- Notes on the above -------+---------+---------+
General notes on NFS issues:
For this upgrade to 64-bit time_t's, the change is so disruptive that I
couldn't get NFS-mounts to work if I booted a "32-bit time_t system"
(ie: 32-bit versions of /bin, /sbin, /lib, ...) on a 64-bit kernel. So,
I added the installworld_oldk script. This script does two things:
1) Creates a mini-/bin inside /boot/kernel.
2) Does a minimal installworld (while still on the old kernel),
thus making it possible for NFS-mounts to work when you reboot.
The first half is a step that would be perfectly safe to do, for any
upgrade (including non-NFS ones), at any time. It is a generally safe
and interesting idea, although it really should be implemented as an
official target in /usr/src/Makefile to be done right.
The second half would USUALLY be a bad idea to do, but I think it's the
only way I can get this specific upgrade to work for people that install
from NFS-mounted directories. It is bad because you are clobbering parts
of your system even though (in the usual case) you would not know that
the new kernel actually works on your system. It also does not do a
full-install, so you end up booting into a system which is part old-
world, and part new-world. It looks like we can get away with that for
this upgrade, but the tactic would be too risky for "standard upgrades".
These instructions assume that you are already familiar with how to do
installations over NFS-mounted partitions. If you are not, you might
want to read other references, such as 'man development'.
Notes on step NFS 8a: ifconfig -a
This shows to the configuration of all your ethernet interfaces. Write
down the IP address and netmask of your main interface. This is
particularly important if the machine obtains its address via DHCP.
You will not be running dhclient after the reboot in step 8, so just
re-use the IP address that the machine is using for the present reboot.
Notes on step NFS 8b: shutdown now
This will drop you into single-user mode, without rebooting. It
will ask if you want to use /bin/sh for your shell. You do.
Notes on step NFS 8d: sh installworld_oldk
Note that this script only installs *part* of the new world. You will
still have to reboot into single-user mode and do the full installworld.
The installworld_oldk script will ask you if you want to build a
mini-/bin. For this upgrade, you should say "yes".
Notes on step NFS 15a:
On my Ultra-10, I have the 'hme0' device as my ethernet card. The output
of 'ifconfig -a' (from step 'NFS 7a') included the lines:
hme0: flags=8843<UP,BROADCAST,RUNNING,SIMPLEX,MULTICAST> mtu 1500
inet 192.168.1.18 netmask 0xffffffe0 broadcast 192.168.1.31
So for this step, I typed in the command:
ifconfig hme0 inet 192.168.1.18 netmask 0xffffff00
Notes on step NFS 15b: mount_nfs
At this step, you may need to specify the host as an IP address instead
of a hostname, because the machine will only be able to resolve hostnames
that are in /etc/hosts.
In my case, I found it easier to create a source file ahead of time
which included the ifconfig and mount_nfs commands that I knew I would
need, and then I just sourced that file after rebooting into single user
mode. If you made such a source file and put it in your root partition,
perhaps under /boot, then that file could also include all of the steps
from 11 through 15c.
Also, it is best use the 'mount_nfs' command, instead of 'mount -t nfs'.
If you use the 'mount' command for NFS mounts, it will turn around and
directly execute /sbin/mount_nfs, and that is not desirable in this case.
Notes on step 17: sh installworld_newk
This script will do some setup work, and then ask you if want it to run
'make installworld'. Most people should just answer "y" (yes) to that
prompt. You can avoid the prompt by including "-y" or "-n" on the
command. If you say "n" (no), then it will tell you what commands
you must type to do the actual installworld.
The script also recognizes a "-S" parameter, which causes it to use
symlinks instead of making copies of programs used by the installation
process. This option will cause less filespace to be used up in /tmp,
but it might be slower in some cases (especially for installs using
an NFS-mounted directory for /usr/obj).
Both this script and the installworld_oldk script also recognize a "-M"
option. This option causes the script to use the absolute minimum PATH
setting that "should" be needed to complete an install. This option is
mainly just for debugging the scripts, though. If you request the
minimum PATH, and some important file was NOT properly copied, then the
installworld will immediately die at that point. This might be painful.
Without "-M", the same oversight would mean that you will run the wrong
*version* of the command, but that older version might actually work
perfectly fine. I did all my testing with "-M" to make sure I had
found all important programs, but there is probably no advantage for
using it for standard system upgrades. Also, if there are no important
files overlooked, then "-M" will not make any difference at all.
# -------+---------+---------+ Upgrading Ports +---------+---------+---------+
Similar to the recommendation for the upgrading the system, I suggest that
you do not 'cvsup' your local copy of the ports collection before trying to
rebuild everything for 64-bit time_t. For one thing, you will have a cvsup
compiled for 32-bTT (32-bit time_t's), and that will not work well on a
system which is using 64-bTT. You might find that you have to 'cvsup' for
some ports, but you will need to get a 64-bTT version of cvsup before you
can do that.
One tactic to use for upgrading ports is to rebuild your already-installed
ports one-at-a-time. If you want to do that, and if you use portupgrade
to upgrade your ports, then I suggest the first thing you should do is:
portupgrade -Rr -f ruby portupgrade #- Ports 1.
Aside: if you get an error about the "ruby-rdoc" port,
then enter: pkg_deinstall ruby-rdoc
and repeat the original command.
portupgrade -Rr -f bash #- Ports 2.
If you have 'bash' installed, or include any other shells
which you have installed from the ports collection. If
your session is *using* one of these shells, then logout
and log back in after recompiling that shell.
portupgrade -Rr -f ezm3 cvsup-without-gui #- Ports 3 (maybe).
If you want to rebuild a 64-bit time_t version of cvsup.
Note: ezm3 (modula-3) needs a patch to work correctly after the
change to 64-BTT. That fix has not been commited to the port
yet [as of Mar 10th], but hopefully it will be commited soon.
There are pre-built packages available for ezm3 and cvsup-without-gui on
the new 64-bTT systems. This ezm3 package *does* include the necessarily
patch. These files are available on the standard ftp servers for FreeBSD.
If you have previous versions installed, then remove them with:
pkg_delete cvsup\*
pkg_delete ezm3\*
If you get warnings about "unable to completely remove" some
lib/m3 directories when deleting ezm3, then also enter:
rm -rf /usr/local/lib/m3
You can install the new packages with:
pkg_add ftp://ftp3.FreeBSD.org/pub/FreeBSD/ports/local-distfiles/gad/ez...
pkg_add ftp://ftp3.FreeBSD.org/pub/FreeBSD/ports/local-distfiles/gad/cv...
Replacing "ez..." with "ezm3-64btt-1.1_1.tbz" and "cv..."
with "cvsup-without-gui-64btt-16.1h.tbz". You can also use
some other standard ftp server, instead of ftp3.FreeBSD.org.
"Now look over all the other ports you have installed, and
re-compile everything that probably needs to be recompiled".
If you are going to do it piecemeal, the next ports to force-rebuild would
probably be languages like perl and python, if you have them installed.
After that, force-rebuild the ports like autoconf and automake, if you
have them installed. Or you might want to play it safe at this point,
and simply recompile *every* port that you have installed.
A different tactic to use for ports is to remove *all* ports before you
do the installkernel/installworld step (while you're still on a 32-bTT
system). Then, once you're up on the 64-bTT system, start making them
one-by-one. If you follow this tactic, you might want to save the output
of a 'pkg_info' command before you start removing ports.
# -------+---------+---------+---------+---------+---------+---------+---------+
If you run into problems when making this change, please report them to
the mailing list freebsd-sparc64@FreeBSD.org .
# -------+---------+---------+---------+---------+---------+---------+---------+
# Notice that the following command can be useful in some settings:
grep '#\- ' UPDATING.64BTT